基于图重构的社交知识推荐  

Social knowledge recommendation based on graph reconstruction

作　　者：张馨月 高辉[1] Zhang Xinyue;Gao Hui(School of Computer Science&Engineering,University of Electronic Science&Technology of China,Chengdu 611731,China)

机构地区：[1]电子科技大学计算机科学与工程学院,成都611731

出　　处：《计算机应用研究》2024年第12期3607-3613,共7页Application Research of Computers

基　　金：四川省科技计划资助项目(2023YFG0021)。

摘　　要：现有推荐模型大多聚焦于显式地构建用户和物品的联系,忽视了对图结构高阶全局特性的建模,对用户隐式兴趣的挖掘不足。因此,提出了一种基于图重构的社交知识推荐模型(social knowledge recommendation based on graph reconstruction,SKRGR),引入图重构技术将用户-物品知识图谱划分为多个子图,并利用基于关系的图注意力网络和三层图神经网络对其进行独立编码。通过实施邻域增强策略,深入挖掘用户的隐式兴趣,促进了对邻域局部特征的建模。为进一步提升节点表示的质量,提出了全局对比学习机制,统一全局交互图和局部协作图的节点表征,并运用门控融合策略控制全局与局部的信息聚合。在Ciao和Epinions社交数据集上的实验结果表明,SKRGR模型在recall指标上平均提升了8.77%,在NDCG指标上平均提升了13.40%,相较于DSL、CLDS等基线模型展现出了显著的性能优势,验证了其在捕捉用户隐式兴趣和建模全局特征方面的有效性。Most existing recommendation models focus on explicitly constructing the relationship between users and items,neglecting to model the high-order global features of graph structure and under-mining the implicit interests of users.Therefore,this paper proposed a social knowledge recommendation based on graph reconstruction(SKRGR),which introduced graph reconstruction technology to decompose the user-item knowledge graph into multiple subgraphs,and utilized relationship-based graph attention networks and three-layer graph neural networks to encode them independently.By implementing the neighborhood enhancement strategy,it deeply explored users’implicit interests and facilitated the modeling of neighborhood local features.To further improve the quality of the node representations,this paper introduced a global contrastive learning mechanism to unify the node representations of the global interaction graph and the local collaboration graph,and used a gate fusion strategy to control the aggregation of global and local information.Experimental results on the Ciao and Epinions social datasets indicate that the SKRGR model achieves an average improvement of 8.77%in recall and 13.40%in NDCG.Compared to baseline models such as DSL and CLDS,SKRGR demonstrates significant performance advantages,validating its effectiveness in capturing users’implicit interests and modeling global features.

关 键 词：社交推荐 知识图谱 对比学习 邻域增强 图神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]